Well, hello everybody and welcome to the Penguin Solutions Fireside Chat at the Goldman Sachs Communacopia + Technology Conference. I have the privilege of hosting Kash Shaikh, CEO of Penguin Solutions. My name is Kat Murphy and I cover Penguin and IT Hardware here at Goldman Sachs. We will have about 35 minutes for today's discussion, inclusive of Q&A towards the end. Maybe to go ahead and get started, September marks seven months since you have joined Penguin as CEO. You joined at an interesting time as the company transitioned from a holdco into an AI solutions-oriented business. For the purposes of this audience, can you help us understand your current business mix and where Penguin is really positioned to go after this investment cycle around AI?
Thanks, Kat. In the last seven months, we have focused on three main areas. First, prioritizing our data center AI infrastructure business and our Integrated Memory business. These two businesses have very high demand because of the super cycle of the infrastructure, and they represent our AI-driven businesses. Second, we have increased our investments in product innovation. Third, we have accelerated go-to-market execution with a clear focus on two main segments: the neocloud segment and the enterprise segment for our AI infrastructure business. We also introduced what we call our AI Factory Platform. This platform combines differentiated products like an OEM offering, as well as a system integrator like end-to-end services, design, build, deploy, and manage. This allows us to be the single partner for our customers, building the factories, AI factories, as well as operating the factories on their behalf. The results show the progress.
In Q3, which was our last earnings that we announced publicly, our AI-driven businesses represented 74% of company net sales, and they grew 104% YoY. The company also had record net sales at company level. Based on our guidance that we provided in the last earnings forecast, we are expecting yet another record quarter at company level, and it will be our first over $500 million net sales for the company.
Great. You also provided a preliminary outlook on earnings last quarter for FY 2027 and guided to 30% YoY growth at the consolidated level, which implies an acceleration sequentially. Can you talk about what is informing your confidence in that sequential acceleration? I know there are some moving parts within the business. Where in the business across those three opportunities that you outlined do you see the most incremental upside opportunity?
As we mentioned in our last earnings, our confidence comes from, first of all, the strength of the market we play in. We have two tailwinds, very high demand data center AI infrastructure market, very high demand memory market, and we have a very unique product market fit for both of these product lines. Combine that with the visibility that we have with our bookings, backlog, as well as the pipeline that is really driving the outlook for FY 2027. In terms of the upside for the outlook, it really depends on how fast we can convert some of the large AI factory opportunities that we have in our AI infrastructure business.
Gotcha. You have reported historically the company in three different segments, but I want to focus on Advanced Compute and then the Integrated Memory opportunity. First, on the Advanced Computing side, can you break down the lineup of hardware, software, and services within your portfolio and how you are going after both the neocloud and enterprise opportunities that you talk to with those products?
Within our Advanced Computing, we primarily focus on our AI infrastructure solution business. This is where we offer the AI Factory Platform, and this AI Factory Platform is a very unique combination of products. These products include ClusterWareAI, which is an AI factory operating system. Just like an operating system like Windows will provide a platform to combine all of the resources between memory as well as applications and CPUs, our AI factory ClusterWare operating system provides a similar consolidation of the cluster, as well as what we have done with our ClusterWareAI recently in the last six months. We have also created agentic experiences. As an example, the operators can now use the natural language to find out the health of the GPUs as well as the utilization of the GPUs. So that is one of the key differentiator from the product perspective in our platform.
We have also invested in our MemoryAI product line. This is a product line where we are leveraging our unique insights from decades of experience in the memory architecture as well as the AI infrastructure data center build-outs. This MemoryAI product line provides benefits especially for the inference workloads. As AI moves from training to inference powering agentic AI, this is where the context sizes of the messages are very long, and these require memory architectures that can help accelerate the LLM responses and performance. So that is the next product line. We have also introduced ComputeAI product line within this portfolio. What it is, essentially, is a set of GPUs and accelerators from NVIDIA as well as from AMD. So we can provide the products to our customers that help them manage the factories as well as build out the factories.
Then we combine them with our end-to-end services. So we get involved much earlier in the cycle with our customers. So we come in, we design the factories on their behalf, and as we are designing the factory, we recommend our products, partner products, and then we also build out, as in do the system integration for them. And then we sign contracts with them three to five years to manage those factories. So this is this unique combination of offering like an OEM product company. Combine that with an end-to-end system integration offering that allows the customer to work with us as a single build partner as well as a single operate partner for the AI factories, primarily focusing on neocloud build-outs as well as large enterprises.
That's very helpful. So you talked about AI factories, MemoryAI, ComputeAI services. That's the appropriate way to summarize them. And you're going after different customer verticals. For the neoclouds, which of those four products resonate most? And for the enterprise, is there a different kind of go-to-market or sales motion that you're trying to engage with your customers on?
It depends on the journey of the customer, right? Whether it is a neocloud customer or an enterprise customer. In some cases, we start small, as in they may be looking for GPUs, and we provide them GPUs with our ComputeAI product line, whether they are enterprise or neocloud, and then we expand with our services or other products. So that's our land-and-expand strategy across both enterprises and neocloud providers. However, in some cases, the land is pretty significant.
What we are seeing, especially in the last six months, what we have seen is the acceleration of the neocloud providers because both the hyperscalers that are giving them the business as well as the business they get from the enterprises, that creates a very unique opportunity for Penguin Solutions, where these neocloud providers are relatively newer providers, so they are setting up the entire data center. So they will go out and, let's say, lease the data center and the power. Then we go in and we do everything from them, as in design it, procure the hardware for them and the software, build it out, and then we are signing contract to help them manage the factory so they can focus on getting the off-takers, as in their customers, and we are the builder and the operator for their AI factories.
It really depends on where they are in the journey. In some cases, it's a smaller land with expand, in some cases, a pretty significant land, and then they continue to expand with us as they are building out new data centers.
Something that's also unique to Penguin is your agnostic approach to other third-party hardware within some of these ecosystems that you're helping set up. Can you talk about how that is a point of differentiation, and why and what in Penguin's portfolio, maybe this goes to the ClusterWare software operating system layer, allows you to bring forward a best-in-breed solution rather than a sole vendor solution?
Yeah. We believe in, first of all, an approach that is much more focused on meeting the objectives of our customers. Let's say it is a neocloud customer. The neocloud customers typically sign the SLAs with their off-takers or their customers. Based on those SLAs, we get engaged, as I mentioned, much earlier in the cycle with them. As we are designing those AI infrastructures for the factories, we build the architecture in a way that it meets their requirement, whether it is our products or it may be the products of our partners such as Dell, along with the ClusterWare. The advantage of the ClusterWare, especially with our engagement with the customers, which starts much earlier than the product companies, we have the discussion and conversation at the architecture layer, as in you have these SLAs.
For you to meet these SLAs, this is the kind of architecture we represent. Also the fact that our ClusterWare is hardware agnostic. Whether it is NVIDIA, whether it is AMD products, we can provide them the support of this operating system that is helping them manage their GPUs irrespective of the vendor, and helping them achieve their objective without necessarily either just dropping the hardware on their doorstep or necessarily forcing them to use a single vendor, which is usually not the case because, as I said, depending on their SLAs, they have different requirements, and we meet the customer where we believe we meet their requirements versus just forcing our product.
Maybe to help illustrate all of the opportunities that you have and the ways you are participating in these types of engagements, could you talk through an example? I know publicly you have talked about a Tier 1 financial customer that you won jointly with Dell. Maybe talk more about what that sales motion looked like, where you are involved versus where Dell is involved, and how you are capturing value across that design, build, deploy, and manage cycle.
Yeah. It is actually a good example of some of the things I have mentioned. So this large bank put out an RFP, and their driver was they were consuming the AI infrastructure from a neocloud provider. However, they had some new applications that they are developing, which they believe that will help them create new revenue streams. So they put out an RFP, and in that RFP, obviously we competed against the product companies, product OEMs, we competed against system integrators. Their requirement was they were building this on-premise AI factory the first time, and they needed someone who not only can provide them the product but help them manage this factory, because these factories are pretty complicated, and it requires deeper understanding of the product and architectures to be able to deliver on the requirements. We were able to win this RFP.
When I sat down with the CIO to understand why us versus. Obviously, we had very large product companies on the other hand, and very large system integration company. It came down to three areas at the high level. First of all, they were looking for something that is much more comprehensive. As in not necessarily just give me this product, this product, and this product. They had their objectives, which is, "I have new revenue opportunities, and I need someone to help me with the factory, but my goal is to develop these applications that will help me with revenue generation opportunities." So the fact that we provide the whole design, build, and deploy to meet their objectives was an advantage, so that they are not piecemealing all of these products to be able to stitch together their first on-premise factory, which is pretty significant size.
That was number one. Number two was our ClusterWareAI product, because again, that ClusterWare makes it easier to manage and automate the monitoring and management of the AI factories. The third one, interestingly, was our MemoryAI product line. They were the first customer who actually bought it, considered it, now we have other customers. They wanted to make sure they were considering the applications that are much more context-rich inference applications, and they wanted to make sure that they have technologies that can enable the performance and faster response for the LLMs. So those were the three differentiators for us in terms of winning that project.
That's very helpful. I want to touch both on the ClusterWare piece as well as the MemoryAI, and maybe transition to the memory side of the story as well. First on ClusterWare, how are you using ClusterWare as part of your sales motion? Maybe said differently, is this something that is included in part of the pitch as to how you can manage an estate across the end-to-end nature of it? Or is it something eventually you could package and sell separately? What are the alternatives to ClusterWare? What are AI factories who aren't using ClusterWare doing as an alternative? Anything to just contextualize the moat that you may have built with that platform.
Yeah. First of all, the current motion of ClusterWare is primarily with our services. As we get involved, whether it is design or management services, ClusterWare is the software our managed services team is using. However, our goal in ClusterWare is to make it standalone, and especially with the agentic experiences that we are creating. These agentic experiences will allow users who are not as much familiar with our product can easily use it, as I mentioned, with natural language interface that we have provided. So short term, primarily a differentiator for us to win larger projects with our services business. Longer term, our goal is to make it available to the customers, whether these are neocloud customers and enterprise customers.
That will also help us more of with the foot in the door land and expand strategy, where we may not have the entire project, but we start with a software, and then as they realize the benefits of the software, we can offer other products and other services. In terms of the alternatives, that's an interesting question because there isn't an equivalent of ClusterWare, because most of the alternative software available, they are very vendor specific. The challenge in the vendor specific software is, again, you have to piecemeal different pieces of infrastructure when you are building the factory or managing the factory. So the challenge becomes increased time, manual configuration, and may not result into the best utilization of the infrastructure.
That's very helpful. Let's talk more about the MemoryAI compute platform or MemoryAI platform, as well as the offering you have within the broader Integrated Memory portfolio. What exactly are you selling within Integrated Memory, and how is MemoryAI an extension of that, and what are your core competencies that are starting to drive differentiation in the way that you serve customers as inference demand increases?
In our Integrated Memory business, we build memory modules for OEMs. These are large OEMs. You can think of large networking companies, large compute companies. Our moat there is essentially, at the high level, engineering expertise to design those memory modules, because this is not just memory, it is a memory card that plugs into as an interface within the OEM product. We have engineering capabilities, we have validation capabilities. We build them, we validate them, and then we have manufacturing capabilities. We have factories where we manufacture those cards. Then we have the supply chain capabilities because we manage this business at scale.
What we have done recently with this business at the high level, first of all, we are primarily prioritizing data center customers, because the memory demand is high everywhere, but a lot of this demand, in some cases, is a function of just supply and demand. We believe the demand in the data center is driven by AI, and we believe the demand in the data center will be more durable than the cyclical memory cycles we have seen, which is why we are very disciplined and focused on the data center. That is one of the things we are changing, and it is helping us both in the short term, and we believe it will help us with the durability of the business. The second thing is focusing on new AI infrastructure opportunities.
Looking at the use cases, such as the inference I mentioned with the adoption of agentic AI, requires more memory in terms of having the capability to store the context so that the GPUs don't have to compute it all over again as the LLMs are doing the inference for an application. CXL is one of the standard Compute Express Link. We have been the early adopter of CXL. Using CXL, we create two products. One of the product is our memory expansion cards. These are just the CXL-based memory expansion cards based on that standard that we provide for the OEMs as a part of our OEM go-to-market with the Integrated Memory business. Then we took that CXL capability and created an appliance, which is MemoryAI KV Cache appliance.
You can think of an appliance which has both the hardware and the software to be able to provide the memory capabilities for the LLMs to store the context. Let us say if the LLM is writing a book, one of the option is every time it is writing a new sentence, it is doing the whole computation again. The challenge becomes inefficient use of GPUs and high bandwidth memory. Or you can have a memory bank, like our MemoryAI KV Cache appliance, where you can store the context of the book you have written so far, so next time when you are writing the next sentence, you don't have to compute all over again, as an example. That is really the benefit. What it really does is providing you the faster response for the LLM, and it is also saving the computation power for the GPUs.
What is the net effect? You may not need GPUs for all the computation, so your spend can go down, and then you will have faster responses for the LLMs, which means better economics and faster responses. That product is part of our AI Factory Platform data center design, build, and manage because it is an appliance that is connected to the GPUs. Whereas the CXL cards are the part of the go-to-market for the Integrated Memory business.
You've talked about the focus on neoclouds and enterprise as customers, but these CXL memory cards seem like they would also resonate with the hyperscale type customer, which would be part of this prioritization of data center. Is that also an opportunity, either from a shipment perspective, a technology licensing perspective? Can you go after the CXL hyperscale part of this up market?
That is definitely an opportunity, and especially with the newer focus of our Integrated Memory business on the data center in general. We are working with data center customers, including hyperscalers that are considering our CXL, especially in the environment where, let's say DDR5 is a bottleneck. They are looking at all the possibilities to be able to get the maximum throughput as well as the performance for their GPUs. Then, as I mentioned, the MemoryAI can be both hyperscalers, can also be neocloud providers, and the enterprise use case with the Tier 1 financials.
Got you. Something unique to Penguin Solutions' model as well with an Integrated Memory is that memory is largely a cost that you get to pass through, and there's also some opportunity in instances to build inventory and earn a spread in markets like this where prices are up significantly. Can you talk to some of the dynamics that are influencing the growth in revenue and the margin expansion opportunity within memory? Maybe by breaking up what is coming from market price increases versus incremental drivers of demand.
Yeah. Overall, our growth in the Integrated Memory business is a combination of both. So prices are going up, which are driving the revenue higher. However, what is more encouraging is the volumes. The volumes are increasing across the board, especially as we discussed for the data center product. While let's say our Integrated Memory business in Q3 grew 111% YoY, we had as much backlog remaining in the bookings that we have received that we are going to be able to ship in the future, and that's the function of the volume growth. And the prices, obviously, at some point may stabilize. What we see is the increase in volume will still drive higher revenue opportunities for us, especially in the data center.
Can you talk to how Penguin is navigating through some of the supply scarcity? Obviously, you're building a backlog, but demand and availability are both informing that. How are you working with suppliers, and is there anything unique about the types of OEMs that you're serving in Integrated Memory that may insulate you from some of the impacts of memory scarcity in the broader market?
Yeah. There is no denying that there are supply challenges in the overall market. However, some of the things that are helping us and are strategic advantages for us, we have a set of OEM customers in the Integrated Memory business that have supply agreements with the memory suppliers, whether it is Micron or SK hynix or Samsung. And that represents about 50% of our business, depending on the particular quarter. Then we have this other segment of the customers where we procure the memory, and this is based on our long-term relationship with the memory suppliers, as well as we have some contracts with the suppliers to be able to acquire the memory because part of the business is not just the memory. We do the design as well as validation and then the supply chain for those parts.
But in the dynamics of, let's say, when we are procuring the memory on the behalf of the customers who don't have the supply agreements, we have more pricing power, right? Because we're doing two things for them. We're procuring the memory for them in a supply-constrained environment, in addition to creating more value with our design and manufacturing of the board. So in general, we feel confident that we will continue to drive the growth in this business. While there are challenges with the supply, we have arrangements that are helping us with the continued growth of the business.
Great. Maybe I'll ask one more, and then we'll open it up to see if there's any questions. Looking beyond the innovations around CXL, which is still in the early days of adoption, you've also talked about developing new architectures in photonic memory. What are some of the projects and work that you're doing there, and when do you expect to see the benefits of photonic memory appliances or memory cards start to impact the opportunity within this segment?
Our work in photonics is based upon our early investments in Celestial AI. Celestial AI was a photonic startup, which was recently acquired about six months ago by Marvell for $5.5 billion. In addition to getting the proceeds from our investment, we continue to work with them on developing an appliance that will allow very high-speed connectivity between GPUs and the HBM, high-bandwidth memory. Because right now, HBMs are primarily directly attached and connected within the GPUs. That kind of interface, as in a photonic interface, will allow the expansion of the high-bandwidth memory beyond the direct-attached memory to the GPUs. As we know, there is nothing faster than the light. So that gives us the opportunity to really create value of a shared memory platform for accelerated applications.
The availability of that product will be second half calendar year 2027, even if we have active R&D work going on with Celestial AI. But it represents the long-term opportunity building upon our CXL memory expansion cards, MemoryAI KV Cache appliance based on CXL. This becomes a natural evolution for us, again, focusing on the data center with our Integrated Memory business and focusing on the next generation architectures and represents a long-term growth opportunity for the company.
Great. Any questions in the audience? I have a couple more here, too. Let's talk about the two segments together and the unique advantage of having both the expertise on the memory side as well as this very services-oriented platform on the compute side. Why is that interconnected nature of the two segments increasingly important for Penguin, and how does it differentiate Penguin against, in the example of the Tier 1 financial win, against some of the other projects that you've bid against?
Yeah. At the high level, first of all, it's a very unique position in the market, especially as AI transitions from training to inference, memory becomes really strategic for the next phase of AI. We believe that allows us to continue to innovate at the intersection of AI infrastructure and memory, continue to create innovations that can help us lead the second phase of AI with inference taking off. At the same time, we plan to continue to invest and differentiate in our AI Factory Platform. The ClusterWareAI that I mentioned, we plan to continue to invest in making it much more agentic and much more easier to use for the customers to build and manage their AI factories.
The continued evolution, continued investment in our AI Factory Platform, also using our unique position at the intersection of memory and AI infrastructure to create new solutions. At the same time, continue to create operational efficiencies so we have the operating leverage and deliver the profits to the business and our shareholders.
We've talked about all of the ways in which you are enabling AI adoption for hyperscalers, neoclouds, and enterprises, but Penguin is an enterprise itself. What applications of agentic AI are you seeing internally, and how is that helping some of your initiatives around things like investing in ClusterWare, go to market as you expand your customer type, things of that nature?
Within our company, we are using AI across all of our functions. One of the use cases that we have realized that is where we are seeing the productivity and the return, it's a pretty clear ROI, is the code generation. We are using agentic AI to develop the code and deliver much faster code for our ClusterWareAI. That's one of the use case. Interestingly, what we see, especially with the large enterprise customers, that is one of the use case they are considering when they are moving or expanding beyond neocloud, consuming AI infrastructure from the neocloud to building their own AI factories. What we see with these enterprise customers, the driver is really economics.
What happens at scale for code generation application, even if they deliver a pretty clear higher productivity, the cost become the challenge because they have to pay for GPU as a service, which is a subscription, and they typically are using frontier models, which are also quite expensive at scale. When they are moving to on-premise, they are obviously spending a lot more upfront with the CapEx, building the factory, but then they don't have to worry about the subscription of the AI GPU as a service, as well as they are considering open weight models increasingly so they can run these high volume persistent inference applications, such as code generation at scale on premise much more effectively with higher security data retention and privacy, along with better economics that they would get in the cloud operating model.
Great. We have about a minute left here, but any final thoughts would be great to know as you look to your first full year as CEO and beyond, where your time is going to be most focused from a strategic priorities perspective.
Going back to the two high-demand businesses, we will continue to focus on capturing our fair market share, especially as AI transitions and expand beyond training to inference in the data center AI infrastructure business, as well as Integrated Memory business. At the same time, continued investment in innovation, whether it is the software, MemoryAI appliances, or the compute appliances. Last but not least, continue to drive up operational efficiencies with using AI across the company to deliver and continue to improve our operating leverage.
In the end, we believe while we have started our journey and the business is growing and we have created value for our shareholders, all of the things that are happening with the infrastructure super cycle of more than $7 trillion that will be invested in the next three years, we believe Penguin Solutions is very uniquely positioned to capture large share of that investment and create long-term value creation opportunity for our investors.
Great. Thank you very much for being here, Kash.
Thank you.
Appreciate it.